A Distributed and Scalable Time Slot Allocation Protocol for Wireless Sensor Networks_Mobile Computing_2011

A Distributed and Scalable Time Slot AllocationProtocol for Wireless Sensor NetworksChih-Kuang Lin, Student Member, IEEE, Vladimir I. Zadorozhny, Member, IEEE,

Prashant V. Krishnamurthy, Member, IEEE, Ho-Hyun Park, and Chan-Gun Lee

Abstract—There are performance deficiencies that hamper the deployment of Wireless Sensor Networks (WSNs) in critical monitoring

applications. Such applications are characterized by considerable network load generated as a result of sensing some characteristics

of the monitored system. Excessive packet collisions lead to packet losses and retransmissions, resulting in significant overhead costs

and latency. In order to address this issue, we introduce a distributed and scalable scheduling access scheme that mitigates high data

loss in data-intensive sensor networks and can also handle some mobility. Our approach alleviates transmission collisions by

employing virtual grids that adopt Latin Squares characteristics to time slot assignments. We show that our algorithm derives conflict-

free time slot allocation schedules without incurring global overhead in scheduling. Furthermore, we verify the effectiveness of our

protocol by simulation experiments. The results demonstrate that our technique can efficiently handle sensor mobility with acceptable

data loss, low packet delay, and low overhead.

Index Terms—Access scheduling, wireless sensor networks, data-intensive applications, topology changes.

Ç

1 INTRODUCTION

SENSOR networks are being deployed for a wide range ofapplications, such as habitat sensing, target tracking, etc.

Wireless Sensor Networks (WSNs), however, perform poorlywhen the applications have high bandwidth needs for datatransmission and stringent delay constraints. Such require-ments are common for Data-Intensive Applications (DIAs). Asan example of a DIA that utilizes a WSN infrastructure,consider the task of Structural Health Monitoring (SHM)system [9], [42] that monitors the integrity of civil andmilitary structures. Wireless sensors observe excitationsaround a surveillance structure, perform data gathering,and periodically report sensed data to a base station (BS).Another example of a DIA is the near-continuous monitor-ing of heat exchangers in a nuclear power plant. Suchapplications generate considerable network load in a shortperiod of time. As a result, collisions between packets canbe a considerable obstacle to achieving the requiredthroughput and delay for such applications. As the dataload increases, we observe severe degradation of networkperformance. Packet success ratio drops due to frequentcollisions and retransmissions. The data glut results in an

increased time delay to reach the sink and an increase in theoverall energy consumption in the WSN. After a certainload threshold, the performance characteristics of tradi-tional WSNs become unacceptable. In addition, wirelesssensor networks may include mobile platforms (e.g., mobilerobots) [23], [24], [25] carrying wireless sensor devices thatcan be deployed in conjunction with stationary sensornodes to acquire and process data for applications, such assurveillance and tracking, environmental monitoring inhighly sensitive areas, or search and rescue operations.Resource constraints in mobile WSNs make it difficult toutilize them for advanced environmental monitoring thatrequires data-intensive collaboration between the sensors(e.g., exchange of multimedia data streams) [19], [26] andcollisions further exacerbate the problem. Also, applicationsmay have stringent requirements on the response time fordelivering the query results. Minimizing sensor queryresponse time and simultaneously minimizing energyconsumption per query is crucial. In general, the time/energy trade-offs involve energy and time gains/lossesassociated with specific layouts of the nodes. Properpositioning (relocation) of mobile sensors has a consider-able impact on the time/energy trade-off [22].

Methods of wireless channel access of sensor networkscan be classified into two categories: random access andscheduling access [28]. Under the random access category,even high-rate wireless networks (e.g., IEEE 802.11) usebest-effort service that can lead to packet loss (fromcollisions), delay, and jitter impacting data-intensive multi-media applications [3]. These problems are aggravated inlow-rate wireless sensor networks, such as IEEE 802.15.4[12]. A previous study [9] reported that successful packetdelivery ratio (PDR) in an 802.15.4 network can drop from95 to 55 percent as the load increases from 1 to 10 packets/sec. Note that it is common for Data-Intensive Sensor

Networks (DISN) to generate 6-8 packets/sec, making theproblem quite significant. Moreover, as the PDR drops,

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 5, APRIL 2011 505

. C.-K. Lin is with the Q2S Centre, Norwegian University of Science andTechnology, O.S. Bragstads plass 2E, NO-7491 Trondheim, Norway.E-mail: [email protected].

. V.I. Zadorozhny and P.V. Krishnamurthy are with the Department ofInformation Science and Telecommunications, School of InformationScience, University of Pittsburgh, 135 N. Bellefield Ave., Pittsburgh,PA 15260. E-mail: {Vladimir, prashant}@sis.pitt.edu.

. H.-H. Park is with the School of Electrical and Electronics Engineering,Chung-Ang University, 221 Heukseuk Dong, Dongjak Gu, Seoul 156-756,South Korea. E-mail: [email protected].

. C.-G. Lee is with the School of Computer Science and Engineering,Chung-Ang University, 221 Heukseuk Dong, Dongjak Gu, Seoul 156-756, South Korea. E-mail: [email protected].

Manuscript received 26 Apr. 2009; revised 1 Jan. 2010; accepted 22 Jan. 2010;published online 27 Aug. 2010.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2009-04-0142.Digital Object Identifier no. 10.1109/TMC.2010.163.

1536-1233/11/$26.00 � 2011 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

sensors may retransmit more in order to increase thelikelihood of delivering information resulting in morecollisions, energy waste, and reduced network lifetime.

The issue of performance degradation in WSNs can beaddressed at different layers of the wireless protocol stack.Our specific focus is the Medium Access Control (MAC) layerwhere fine-grained control can reduce collisions and energywaste [11]. Carrier Sense Multiple Access/Collision Avoid-ance (CSMA/CA) is the popular contention-based MACscheme adopted in both IEEE 802.11 and IEEE 802.15.4standards. However, as already mentioned, the perfor-mance of CSMA/CA degrades as the load increases [9],[12]. Scheduling-based access avoids this problem (forexample, Distributed randomized TDMA scheduling(DRAND) [4] and Data Transmission Algebra (DTA) [20]).DRAND provides reliable data transmission and reducescollisions, but the concerns that arise include increasedcontrol overhead and scalability with network size [21] andespecially mobility. DTA efficiently manages data deliveryin DISNs, but it is centralized and its shortcomings aresimilar to those of DRAND. We have previously proposedhybrid access approaches that combine advantages ofrandom access and scheduling access [21]. In this work,we proposed a Cyclic Probabilistic Transmission protocol(CPT) that allows sensors to transmit with probabilities intime slots and a Neighbor-Aware Probabilistic Transmis-sion protocol (NAPT) that controls these probabilities basedon neighborhood information. We demonstrated that CPTand NAPT considerably outperform random access proto-cols (such as CSMA/CA) and are compared with schedul-ing access protocols (such as DRAND). Table 1 provides aqualitative summary of various MAC schemes for WSNs interms of the scalability of access method and expectedutility (throughput without losing packets) in Data-Intensive Sensor Networks.

In this paper, we propose a decentralized techniquecalled Grid-based Latin Squares Scheduling Access (GLASS) thatmaintains graceful performance degradation in DISNs as thedata load increases. Meanwhile, the protocol is designed tobe lightweight, overhead-efficient, highly scalable, androbust in the presence of mobility. Our motivation for this

work stems from our previous research in hybrid accessusing NAPT and CPT. At the one end of the spectrum (seeTable 1), we have distributed but scalable CSMA/CA andCPT schemes with performance limitations due to collisionsduring periods of high data loads. At the other end of thespectrum, we can ensure that there are no conflictingtransmissions in the network using scheduling (e.g.,DRAND and DTA). While these scheduling approachessave on the energy waste due to collisions, they can increasethe energy waste due to overhead from control messages.This concern further worsens as network scale expands,which is likely in WSNs. Schedules also need recomputationif the topology drastically changes due to even somemobility. At the middle of the spectrum, i.e., NAPT, theoverhead concern is alleviated, but its performance does notmatch the performance of scheduling access.

In this paper, we propose a novel approach calledGLASS that provides a decentralized scheme for creatingconflict-free (resulting in high network utility) time slotschedules, analogous to DRAND, as shown in Table 1,while, at the same time, reducing the control overhead.GLASS exploits several opportunities for minimizing over-head cost while maintaining conflict-free concurrent trans-missions in DISNs. In particular, sensors use locationinformation to improve channel access. First, sensorsvirtually divide the monitoring area into a grid. Then, eachsensor associates itself with one virtual grid cell. Thisdesign allows neighboring sensors to maintain spatial andtemporal separation between potentially colliding packetswhile keeping channel access distributed. The novelty inGLASS is its use of the Latin Squares (LS) characteristics[13] to facilitate the assignment of time slots for transmis-sions among sensors within a grid cell, thus, reducing thenumber of colliding transmissions. We demonstrate thefeasibility of our technique using analysis and simulations.In particular, the simulation results show that the controloverhead cost of GLASS is about 70 percent less comparedto DRAND and the performance of GLASS is robust withrespect to changes in topology, while its transmissionefficiencies are comparable to those of DRAND in staticnetworks. Note that GLASS assumes that each node in aWSN is aware of its geographic location.1 While location-based approaches have been adopted in routing mechan-isms for sensor networks [1], [2], to the best of ourknowledge, they have not been utilized for optimizingchannel access of sensors.

The paper is organized as follows: Section 2 considersrelated research. In Sections 3 and 4, we describe the GLASSapproach and provide analysis of the same. Section 5describes the simulation-based study. We offer conclusionsand discuss future work in Section 6.

2 RELATED WORK

Channel access in sensor networks can be classified intoscheduling-based and random access categories [28]. Below,we briefly describe some prior work in both categories. For

506 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 5, APRIL 2011

TABLE 1General Framework of MAC Schemes Classification

1. We note that using global positioning system (GPS) is not alwayspossible in WSNs because of energy- and location-precision constraints.WSNs commonly utilize ad hoc localization methods calculating coordi-nates of sensors using special beacon nodes whose positions are known.Further consideration of this subject is beyond the scope of this paper.

DISNs, which need to support continuous and/or periodictraffic loads, it is more appropriate to employ the schedul-ing approach to manage channel access [28]. The schedulingapproach is mostly adopted into a structure, Time-DivisionMultiple Access (TDMA).

2.1 Time-Division/Scheduling Schemes

We describe some of the time-division-based MAC proto-cols for sensor networks here.

We have previously proposed an algebraic optimizationframework called DTA that performs collision-aware queryscheduling in Wireless Sensor Networks [20]. The DTAapproach makes use of a set of operations that taketransmissions between sensors as input and produce aschedule of transmissions as output by a DTA optimizer. Thegenerated transmission schedule is collision-free due to theknowledge of the collision domains of elementary transmis-sions. We have shown that the DTA framework considerablyoutperforms the existing 802.15.4 CSMA-CA as it enablesconcurrent transmissions to a high degree and reducescollisions [20]. However, the overhead cost of generating theDTA transmission schedule is a concern. This problem canbe further magnified in dynamic networks.

Sensor MAC (S-MAC) [29] is a static-scheduling-basedenergy saving protocol that allows neighboring nodes tosleep for long periods and wake up, both in a synchronizedfashion, to avoid wasting energy from idle listening,collisions, and retransmissions. Thus, neighbors conserveenergy when a node is transmitting. However, S-MAC doesnot provide an on-demand interaction with the receiver (ituses a static sleep interval). Sivalingam et al. [30] proposedan Energy-Conserving Medium Access Control protocol(EC-MAC) for ad hoc networks. Using this protocol, acentral controller is responsible for reservation and sche-duling strategies. The EC-MAC protocol can only operate inan environment where every sensor hears each other.Lightweight MAC (LMAC) [18] implements a distributedtime slot scheduling algorithm for collision-free commu-nications. Time is divided into slots and sensor nodesbroadcast information about time slots, which, as theybelieve, they control. Neighboring sensor nodes will avoidpicking up those slots and choose other slots to control.Within its time slot, a sensor node will transmit a messagewith two parts: control and data. The control part includessufficient information for neighbors to derive a time slotschedule of local sensors so that transmissions amongneighboring sensors will not collide. Sensors must listen tothe control parts of their neighbors. Time slots can bereused at distances where interference is small (three hopsfor instance). With such an algorithm, the goal of collisionavoidance is achieved at the price of extra control overheadand listening time. Chatterjea et al. enhanced LMAC withAdaptive, Information-Centric, and Lightweight MAC (AI-LMAC) [31] that uses captured local data about trafficpatterns to modify operations accordingly. While AI-LMACis adaptive and information-centric, it still shares LMAC’sextra control overhead and faces possible performancedeficiency from unexpected bursty traffic. Both LMAC andAI-LMAC were designed not with the goal of supportinghigh data loads, but with the objective of reducing theswitching time/cost from sleep mode to transmit mode.

Rhee et al. [4] propose a distributed randomized timeslot scheduling algorithm, DRAND that is used within a

MAC protocol called Zebra-MAC [5] to improve perfor-mance in sensor networks by combining the strength ofscheduled access during high loads and random accessduring low loads. The distributed implementation ofDRAND allows a sensor to select a time slot, which isdistinct from time slots of its two-hop neighboring sensors.This feature reduces data packet collisions. The DRANDalgorithm includes two major phases: Neighbor Discovery-Hello and DRAND Slot Assignment. In the neighbordiscovery phase, sensors broadcast Hello messages periodi-cally to announce their existence. Next, sensors exchangecontrol messages like Request, Grant, Release, or Reject todetermine the time slots of sensors. With this scheme, themessage complexity is O(n), where n is the maximum sizeof a two-hop neighborhood in a wireless network. WhileDRAND provides reliable data transmissions, some con-straints are noted. First, this algorithm is suitable for awireless network where most nodes do not move. If thetopology changes dynamically, the algorithm should be runfrequently to ensure delivery reliability. The frequentexecutions of the algorithm will likely consume moreresources in sensors. Next, the algorithm ensures datadelivery by assigning collision-free time slots to sensornodes. However, transmissions can still collide in theDRAND slot assignment phase because of randomizedtransmissions and channel contentions.

Wave scheduling, proposed by Trigoni et al. [34],minimizes packet collisions by scheduling groups of trans-missions (waves) in sensor networks. Another type ofdistributed TDMA protocol was proposed in [32]. Its goal isto permit mobile sensors to move, and then, reallocatethemselves a time slot without involving the entire network.Ali et al. [33] proposed distributed and adaptive TDMAalgorithms for multihop mobile networks. One concern withthis design is that dynamic topology changes may lead tofrequent exchanges of control packets that could consumebandwidth and energy resources. The protocols in [38], [40]combine Latin Squares and TDMA scheduling design. Bao[38] integrates the value of the symbol in the Latin Square intopriority-based channel access mechanisms where a nodewith the highest priority number is granted access to thechannel for one time slot or a node defers the channel accessby the priority number of time slots. Ju and Li [40] propose atime-based scheduling design with Latin Squares in multi-channel networks. While aiming at collision-free channelaccess at any moment, they have limited scalability orfeasibility due to expensive multichannel components forlightweight sensors. A hybrid protocol, NAPT, was proposedin [21]. It combines the features of randomized transmissionand heuristic scheduling where the scheduling feature isused to avoid possible packet collisions and sensors facingfewer contentions apply randomized transmission. It resultsin alleviation of excessive overhead messages, but it cannotguarantee collision-free transmissions.

We have considered the methods deriving time slotschedules at the MAC layer. Next, we review the relatedwork in TDMA-based bandwidth reservation for QoSrouting. Jawhar and Wu [39] address the issue of raceconditions and parallel reservation conflicts in ad hocnetworks. They propose an algorithm based on maintainingthree-state slot status (free/allocated/reserved) at eachnode. It performs synchronous and asynchronous time slotupdates to one-hop and two-hop neighbors and implements

LIN ET AL.: A DISTRIBUTED AND SCALABLE TIME SLOT ALLOCATION PROTOCOL FOR WIRELESS SENSOR NETWORKS 507

smart timers to optimize the throughput and efficiency ofthe selected path. This technique is similar to the GLASSprotocol in the way how conflict-free time slots are assignedamong neighbors. Gerasimov and Simon [41] use a similarconcept and propose a protocol that combines informationfrom the network and data link layers. The authorsmodified the AODV protocol to perform bandwidthreservation, i.e., time-slotted data link control, therebysupporting QoS path reservation and release. They, how-ever, did not consider control phases of sensors, clusterdefinition, and evaluation of control messages.

While the above scheduling schemes are able to reducetime and/or energy waste from possible transmissioncollisions, they simultaneously introduce one or more ofthe following problems. Many of them have considerablecontrol message overhead (e.g., LMAC, AI-LMAC, andDRAND) for building data delivery schedules. In addition,a sensor network is an error-prone and dynamic environ-ment so it may often be difficult to use deterministicschedules (e.g., DTA). For instance, packet retransmissionfor dropped packets (due to channel conditions) may leadto high delays, especially in a large multihop network [20],[27]. In addition, generating and delivering an optimalschedule for data transmissions is complex in a large-scalesensor network [27]. These concerns bring up questionsabout the scalability of the existing approaches.

2.2 Random Access Schemes

Random access techniques implement highly scalable andlightweight distributed medium access control schemes.The IEEE 802.15.4 standard utilizes random contentionaccess using CSMA/CA but it suffers from poor perfor-mance in DISNs [9], [12]. Schurgers et al. [36] proposed acontention-based protocol called Sparse Topology andEnergy Management (STEM) to save energy. STEM imple-ments a two-radio architecture that allows the data channelto sleep until communication is required. Channel monitor-ing alleviates collisions and retransmission. However, abusy tone has to wake up the entire neighborhood of a nodesince the intended receiver’s identity is not included on themonitoring channel. Thus, neighbors waste extra energy.Vaidya and Miller [35] tried to address such issues notconsidered in STEM by introducing a rate estimation (RE)scheme on top of it. They wake up the node, which hadpreviously been involved in communication, via RE usingan optimal wakeup interval. In this way, they minimizeoverall energy consumption. Both STEM and RE schemesassume a two-radio architecture. They ignore the complex-ity of adopting the second radio in a resource-constrainedsensor that may result in difficult transceiver design andadditional energy consumption by the second radio. Notethat a probabilistic access approach [27] can reduce over-head in terms of control messages to access the channel.

3 GLASS PROTOCOL DESCRIPTION

3.1 System Model

We assume the following system model:

1. Initially, sensors are evenly deployed in a field.Each sensor is aware of geographical data (e.g.,coverage and location) associated with the mon-itored area. Sensors can determine their locations

with different accuracy (i.e., with finer of coarserlocation granularity).

2. Wireless sensor networks may include both station-ary and mobile sensors. The mobile sensors, e.g.,mobile robots using sensors for critical monitoringapplications [42], may change locations to performtheir missions. Such sensor relocations are oftensporadic and unpredictable, which makes channelaccess even more problematic for the sensorsdelivering collected data to next-hop neighbors. Weadopt a quasi-periodic traffic model as more appro-priate for DIAs. This model also presents a challen-ging scenario in terms of channel access due to thelarge numbers of concurrent data transmissionswithin a short period of time. Collisions occur whena sensor receives more than one transmissionsimultaneously from different parties.

3. Sensors are uniquely identified and equipped withomni-directional antennas. Every sensor transmits orreceives on a common carrier frequency. It transmitsin assigned time slots and “idle listens” or receivesotherwise. A sensor can transmit a message at thebeginning of a time slot and complete the transmis-sion, including receiving an ACK message, withinthe same time slot (which is 7 ms given a 250 kbpsdata rate). All sensors transmit at the same powerthat covers a limited footprint,2 so multihop com-munication between a source and a sink is antici-pated. Efficient deployment of data-intensivemultihop WSN is challenging. Meanwhile, it isexpected that multihop WSNs will be utilized asan implementation platform for a wide class ofmission-critical monitoring applications with highbandwidth needs (e.g., bridge monitoring anddisaster management). A major reason is relativelylow deployment cost compared to wired alterna-tives. We aim at supporting DIA in such a network.

4. The previous part of GLASS does not need any timesynchronization to run, but the TDMA part using theschedule requires it. Here, time synchronization ismanaged by a Base Station using beacons. Note thatsensors may lose beacon messages and there may beclock drift between synchronizations. Hence, sensorsare designed not to transmit during the firstmillisecond of every time slot. This provides a slacktime for synchronization errors and prevents colli-sions at the boundary of time slots. It tradesbandwidth for the error elimination that can beadjusted using hardware.

3.2 GLASS Protocol

The GLASS protocol includes three phases: Grid Searching(GS), Transmission Frame (TF) Assignment, and Time SlotAssignment. Sensors follow these steps to determine timeslot schedules cooperatively. The protocol is completely


2. Sensors transmit at the same power, but this could result intransmission ranges with variations because of physical environment. Tominimize this problem, the GLASS protocol employs the followingmethods: first, a sensor uses the local neighborhood discovery to learnradio connectivity of its neighboring nodes that helps prevent possibleconflicts caused by overlapping time slot assignments. Next, the config-urations of virtual grid system and CAIG function are adopted to alleviatethe chance of transmission collisions introduced by sensor interference.Their details are described in the following sections.

decentralized. To account for relocation of nodes in amobile WSN, the GLASS protocol either periodically checksthe accuracy of time slot assignment or requires the mobilenodes to verify/update their time slots following the CAIGfunction, as described in the following session. The cost oftime slot reassignment is low (see experimental evaluationin Section 5) due to the low complexity and small overhead.Also, the performance of GLASS is robust with respect tochanges in topology.

3.2.1 Grid Searching

We devise a GS algorithm to assign sensors in a monitoringarea to grid cells. We assume that the monitoring area isvirtually split into square grid cells with uniform shapesand sizes (see Fig. 1). R is the length of one edge of any gridcell, and it ranges between 2r and 2.1r, where r is thesensor’s transmission range

Such a configuration for R is critical for alleviation ofcollisions and overhead. In this paper, we set the R value as2.1r. We will explain this further in Section 4. In addition,each grid cell is identified by a unique ID, associated with itslocation, i.e., a pair of coordinates (GS_Xi, GS_Yi). GS_Xi andGS_Yi represent the vertical and horizontal coordinates of agrid cell correspondingly. Every sensor applies the GSalgorithm, presented in Fig. 2, to determine the ID of the gridcell using its location information. Here, (xi, yi) is thelocation of Sensor i and (X0;Y0) defines the area covered bythe WSN. Using spatial relationships between the sensor andthe monitoring area, the sensor independently calculates theID of the grid cell. If a sensor is located on a border between

two grid cells, namely, the value of Z is an integer (here, Zcorresponds to the x and y coordinates of a sensor in units ofR, the length of a side of the grid cell—see Fig. 2), this sensorwill randomly choose a grid cell’s ID between these two gridcells. Note that we here introduce a simple grid cell searchmethod based on a proprietary two-dimension map. The GSalgorithm will be modified if the geographical coordinatesystem changes. Further consideration of this subject isbeyond of the scope of this work.

3.2.2 Transmission Frame Assignment

After a sensor locates its grid cell, it proceeds with TFassignment. We define a TF as a group of continuous timeslots. The TF structure repeats to handle sensors’ transmit,idle, or receive states. The TF can be divided into multipleequal Subtransmission Frames (STFs) that are orthogonal. Inthis paper, we use a configuration with two STFs: A and B.Fig. 3 describes the algorithm for assigning an STF to asensor. The sensor uses the GS result from the first phase toindependently assign itself an STF (either A or B). As aresult, sensors in adjacent grid cells operate at differentSTFs, reducing the potential for collisions. We can alsoconfigure four STFs via a minor modification of thealgorithm in Fig. 3.

Fig. 4 presents examples of STF assignments. There is anetwork with nine grid cells and five nodes with twodifferent STF scenarios (see Figs. 4a and 4b). Each node usesthe algorithm to find its STF as shown in Fig. 4c. Thus,collisions of transmissions, for example, from Sensors 2 and3 are avoided due to temporal separation. With the fourSTFs configuration (see Fig. 4b), this scheme extendssensors’ transmissions over more STFs and intuitivelyguarantees collision-free intergrid transmissions in thenetwork. However, it may result in worse bandwidthutilization and higher packet delay compared to the twoSTFs configuration because of the reduction in concurrenttransmissions by sensors. Therefore, we pick the two STFsconfiguration for the rest of this paper.

In the two STFs case, the length of a TF is configureddifferently for different sensor distributions:

Length of TF ¼ 2 � STF

¼ 2 � Number of deployed sensors

Number of grids in the network

� �þ �

� �:

The number of deployed sensors and number of gridcells are known in the predeployment stage. � is anadjustable variable between zero and Number of deployedsensors/Number of grid cells. If sensors in the network areevenly distributed, � will be small. If the sensors are not


Fig. 1. Virtual grid network.

Fig. 2. Pseudocode of grid cell search.

Fig. 3. Pseudocode of assigning the subtransmission frame.

evenly distributed, � will increase. To avoid uncertainty insensor deployment, � should not be too small, but this mayslightly impact packet delay due to longer transmissioncycle/frames.

3.2.3 Time Slots Assignment

After sensors discover their GS and STF, the next step is todetermine a time slot for the transmission state of eachsensor. We use LS to assign time slots for sensors, therebyavoiding collisions between neighbors. First, each sensorperforms neighborhood discovery to prepare for time slotsscheduling. The neighborhood discovery requires allsensors to broadcast their information about GS and STFto one-hop neighbors. In this way, every sensor is aware ofits neighbors and maintains a neighbor table which recordsneighbors’ ID, distance/hop count, GS, and STF. Further-more, sensors need to keep complete and accurateneighbors’ information within their grid cells (local data),so each sensor must broadcast newly received data andupdate its neighbor table. Sensors adopt the data aggrega-tion, CSMA broadcast, and ACK techniques to avoid fine-grained and failed broadcasts. Because the side length of allgrid cells is 2.1r, the maximum distance for a sensor toconvey data within a grid cell is three hops. In other words,the sensor needs two or three broadcast messages toannounce itself and discover neighbors’ presence within agrid cell if none of the messages is lost.

Next, sensors utilize the given information about theSTF to generate a Latin Squares Matrix (LSM) for time slotsassignment. An LSM for m time slots of an STF is anm�m array, where each cell of the array contains one of aset of m symbols. Each symbol occurs only once in eachrow and once in each column [13]. An LSM withimmediate sequential symbols can easily be constructed.One method of building a 2k� 2k LSM is explained nextand illustrated in Fig. 5 [14].

1. Enumerate the 2k symbols successively from 1 to n.2. Assign successively the integers from 1 to n to the

n cells in the first row by proceeding from left toright entering only cells in odd-numbered columns,then reversing direction and filling in the cells ineven-numbered columns.

3. In each column, starting with the number alreadyentered in the top cell, proceed downward, enteringin each cell the integer immediately following theone in the cell above it, except the integer n followedby 1.

This way sensors can build an LSM of any size. Wedefine rows of the LSM to represent each local sensor’stransmission and columns of LSM to represent time slots ofthe STF (see Fig. 5). Each sensor selects the local sensors ofits LSM by choosing the sensors with the same GS and STFin the neighbor table. The selected local sensors are sortedby ascending order of their IDs in the LSM rows. WhenValue of a cell of Latin Squares Matrix VLS (i, j) is equal to 1,the corresponding sensor i of the cell is allowed to transmitin its corresponding time slot j. For example, Sensor 5 is setto transmit its data within time slot 5 according to the LSMshown in Fig. 5. The rest of the time slots are either inreceive or in idle states. Note that the sensors can configurethe time slots to the receive state using the information ofone-hop neighbors’ time slots assignment. The rest of thetime slots are in the sleep state for energy savings. Afterthe time slot of a sensor is determined using the LSM, thesensor broadcasts the information of its time slot andneighbor table to one-hop neighbors using CSMA and ACK.

There is no guarantee that the generated LSM rows canbe completely assigned to the local sensors. When thenumber of local sensors is less than the size of LSM, somesensors, e.g., sensor 1 in Fig. 5, are asked to access thechannel more frequently in order to use bandwidthefficiently, but channel access fairness becomes an issue.We adopt the rotating procedure that arranges local sensorsin extra rows of the LSM to give fair channel access to all thesensors inside the grid cell. If the number of local sensors ismore than the size of the LSM, some sensors’ operations aretemporarily omitted based on the rotating procedure. Sucha situation reveals an overcongested grid cell. In otherwords, network resources are unfairly distributed. With theabove procedures, local sensors can create an identical LSMthat ensures the same transmission schedule is shared by allthe sensors in the grid cell if the sensors perform their


Fig. 4. (a) Network with two STFs configuration. (b) Network with fourSTFs configuration. (c) Comparison of STF examples.

Fig. 5. Example of LSM.

neighborhood discovery correctly. The whole process isindependent and distributed for every sensor.

With the LS function, we enable nonoverlapping timeslots assignment inside each grid cell. However, it is stillpossible that collisions will occur between transmissionsnear the intersection of four grid cells because sensors fromdifferent grid cells may use an identical STF. To address thischallenge, we devised a function called Collision Avoidancenear Intersection of Grids (CAIG) illustrated in Fig. 6. Thisfunction is initiated when a sensor with even GS_Yi findsthat its time slot overlaps with the time slots of its two-hopsneighboring sensors. In this case, the sensor updates its timeslot and broadcasts the new neighbor table using CSMAand ACK. The sensors with even GS_Yi are chosen toperform the update since collisions occur near intersectionsof the grid cells with the same STF. Thus, it is sufficient toavoid all of the collisions by monitoring partitions of thenetwork. This design indirectly reduces computation ofthe reassignment of conflicting time slots. Regarding theprocess of updating a conflicting time slot, the sensorchooses the smallest available time slot from the time slotsnot occupied by the sensor and its neighbors within two-hop, so it reduces the chance of exceeding the length of STF.If the length of STF is too short to include the requiredamount of time slots, some sensors will be forced to turntheir transmission states off, resulting in a gap in the WSN.We can avoid this problem by adjusting the � parameter.

We use an example to illustrate the CAIG function.Given the network shown in Fig. 7, A and B denote the STF,sensors are denoted by the letters in circles, and anunderlined red number next to the sensor is the time slotof the sensor (derived from the LSM technique). In the casestudy, we assume that every sensor already knows theinformation about its neighbors in the same grid cellbecause of local neighborhood discovery. A sensor broad-casts the updated neighbor table after determining a timeslot using LSM, so that the sensors located at the boundaryof other grid cells can discover potential collisions. Forinstance, nodes S1 and S7 are one-hop apart and located indifferent grid cells. After node S7 broadcasts its neighbor

data, node S1 will know that they share the same time slot.Node S1 thus starts the reassignment of its time slot (timeslot 5), and then, broadcasts the updated neighbor data.Note that the assignment of a new time slot involvessearching for the smallest available time slot in the sensor’stwo-hop neighborhood, i.e., the two-hop graph coloringproblem [4], [16]. For Node S1, time slots 1, 2, 3, and 4 areused by nodes S7, S8, S2, and S3, respectively. So, the nextsmallest time slot is time slot 5. Similarly, node S4 learnsabout the collision between itself and node S8. Therefore,node S4 locally reassigns and broadcasts its new time slot(time slot 6). An instance of the hidden terminal problem isbetween nodes S2 and S5. They belong to different grid cellsand are two-hops apart. So, the broadcast message from oneis not received by the other. However, when node S3 hearsthe broadcast message of node S5, it finds that its neighbornode S2 shares the same time slot 3 with another neighbor,node S5. Node S3 thus broadcasts its updated neighbor datato help node S2 know of the possible collision. Node S2reassigns its time slot to time slot 2 and broadcasts it. Thus,CAIG alleviates possible collisions by providing distincttime slots to sensors in any two-hop neighborhood.

After the time slot assignment schedule is set (supportedby LSM and CAIG), sensors run the transmission, receive,or sleep mode in time slots by following this schedule. Instatic networks, transmission of data should be conflict-freeand needs no ACK message, but this protocol still adoptsthe ACK function in data delivery for higher layersmanagement, e.g., the regeneration of time slots allocationfor excessive colliding packets.

In this paper, the spatial reuse of channel assignment isbased on TDMA on a single frequency. The channel (timeslot) assignment of GLASS can be converted into afrequency channel or an orthogonal spread spectrum codewith no significant changes whatsoever. We note here that iftime slots are replaced by a set of frequency channels ororthogonal codes, this will imply that the entire transmis-sion occurs during the same time and does not require timesynchronization.

4 GLASS ANALYSIS

In this section, we analyze the correctness and the overheadcomplexity of the protocol.

4.1 Correctness of the Time Slots Schedule

To test the validity of the time slots schedule generated bythe GLASS, we first define what a conflicting time slot is.


Fig. 6. Pseudocode of CAIG function.

Fig. 7. Illustration of CAIG functionality.

Based on the two-hop graph coloring problem, there is aconflict with the time slot, when the time slot of a sensor isthe same as the time slot of another sensor within two-hopsneighborhood [4], [16]. If we satisfy the condition that noconflicting time slots exist between any two sensors from anyone grid cell and from any two different grid cells, we willprove that the generated time slot schedule is conflict-free.

Theorem 4.1. There is no conflicting time slot assignmentbetween any two sensors within any grid cell when theprotocol converges.

Proof. To prove Theorem 4.1, it is sufficient to prove that notwo sensors within two-hop of each other share the sametime slot in any grid cell. Given the side length of a gridcell as 2.1r, the maximum distance between any twopoints within a grid cell is Diagonal (D), 2.98r

D ¼ 2:1�20:5r � 2:98r:

ut

To cover such a distance, three-hop transmission isrequired. Our neighborhood discovery scheme covers theneighbors up to three-hops away within the grid cell. Thisensures that every sensor discovers all of its neighbors withinits grid cell for the correct time slots assignment. Next, weapply the LS characteristic to guarantee distinctive time slotsfor all the sensors located within a given grid cell. Under theassumption of correct neighborhood discovery, local sensorsuse the LS generation algorithms [13], [14] to create anidentical LSM. Such an LSM ensures that every local sensorhas the same time slots assignment schedule as every othersensor, which prevents overlapping time slots within a gridcell (see Fig. 5). If a sensor’s time slot conflicts with itsneighbor’s time slot, then the CAIG function is initiated. Thissituation will not impact the accuracy of the time slotsassignment in any two-hop neighborhood because the CAIGfollows the two-hop graph coloring. Thus, we guarantee thatany sensor selects a time slot different from the time slots ofits two-hop neighbors in the grid cell.

Theorem 4.2. There is no conflicting time slot assignmentbetween any two sensors from any two different grid cells afterthe protocol converges.

Proof. To prove Theorem 4.2, it is enough to prove twopoints: 1) there are no conflicting time slot assignmentsbetween any two sensors that belong to different gridcells and use different STFs and 2) there are noconflicting time slot assignments between any twosensors that belong to different grid cells but use thesame STF. We refer these to the fact that any twoneighboring grid cells share either an identical STF ordifferent STFs in this GLASS protocol. To prove point 1,in the Transmission Frame Assignment phase, weinterleave STFs of nonadjacent grid cells, as shown inFigs. 4a and 4b so that transmissions from the grid cellswith different STFs never overlap because of temporalseparation. This validates the first point. Concerning thepoint 2, transmissions can conflict due to overlappingtime slots either at a common receiver in between (ahidden terminal problem) or around intersections of anyfour grid cells (diagonally neighboring grid cells). In case

of the two STFs setup, the distance between any two gridcells with identical STF is the side length of a grid cell,2.1r, except at the intersections of four grid cells. Such adistance prohibits the sensors from these two grid cells tocommunicate directly if the sensors are not in theintersection region. The sensors are also unable to finda common receiver to relay data because the distancebetween them is greater then two-hops (2.1r). Therefore,there is no common receiver that uses overlapping timeslots to receive data from both sensors. The CAIGfunction is devised to address the conflicting time slotin the cell intersections. Using collected neighborinformation from other grid cells, a sensor near theborder of its grid cell can examine the assigned time slotsto exclude potential collisions. It reassigns itself to adistinctive time slot that is different from the time slotsused by its two-hop neighborhood, if a conflicting timeslot is discovered. As a result, the concerns with thehidden terminal problem and the collision aroundintersections of any four grid cells are alleviated. Thus,we guarantee that there is no conflicting transmissionbetween any two sensors from any two grid cells. tu

Theorem 4.3. There is no conflicting time slot assignmentbetween any two sensors when the protocol converges.

Proof. According to Theorems 4.1 and 4.2, there is noconflicting time slot assignment within any grid cell aswell as between any two grid cells. So, the time slotsassignment schedule with the two STFs configurationproves to be conflict-free. Besides, the time slot assign-ment schedule with the four STFs configuration is alsoconflict-free (as it follows from Theorem 4.1 and point 1of Theorem 4.2). tu

These results regarding the behavior of GLASS are basedon the spatial and temporal relationships between differentsensors. In this section, we do not consider the effects ofradio propagation or interference. We do incorporate thetwo-hop graph coloring problem into the design of GLASSto support conflict-free time slot scheduling. The hopsindicate radio connectivity rather than any particular radiopropagation model.

4.2 Overhead Complexity

The overhead of GLASS occurs during neighborhooddiscovery and the message complexity depends on the areaof local neighborhood (grid cell). Thus, the value of R hassignificant impact on the overhead cost.R ranges between 2rand 2.1r. We set R to be larger than 2r to alleviate the hiddenterminal problem (see the proof of Theorem 4.2) and lessthan 2.1r to minimize the overhead messages by limiting themaximum hop count (i.e., three hops, inside a grid cell, seethe proof of Theorem 4.1). In other words, we configure R toenhance the transmission reliability while minimizing theneighborhood discovery cost. Intuitively, the smaller thevalue of R, the better is the spatial reuse of all sensors.Meanwhile, we take into account the effect of radioirregularity [15] between transmissions, so we keep a bufferdistance to reduce this effect. We set the default ofR as 2.1r topreserve the buffer, i.e., 10 percent of r, avoiding transmissioncollisions while taking into account the spatial reuse factor,the transmission reliability, and the overhead cost.


After neighborhood discovery is completed, each sensorderives the time slots schedule in a distributed way withone or more broadcast messages. These depend on thedegree of local contention. The complexity remains O(x),where x is the size of locally contending neighbors.

5 SIMULATION RESULTS

We implemented GLASS in Network Simulator (ns-2) [10]and evaluated transmission efficiency, overhead complexity,latency, impact of changing topology, and scalability. Wecompare the GLASS protocol with the IEEE 802.15.4 CAPmode (CSMA/CA) [6] and DRAND [4]. We used thecomparison with 802.15.4 as a baseline of this study, andDRAND served as an example of an advanced access-scheduling approach. DTA [20] and NAPT [21] are notincluded in this comparison study since DTA is not adistributed scheme and NAPT does not perform as well asDRAND in terms of transmission efficiency. We set thechannel data rate to 250 Kbps and the sensor transmissionrange to 15 meters. The packet size is 70 bytes. Four networktopologies are tested in simulation, as shown in Fig. 8. Thefirst topology (Fig. 8a) is a random dense network. It includes20 sensors randomly placed in a 63� 63 m2 flat area with aBS, and all sensors intend to report sensed data to the BS. Weplace the BS in two different locations to demonstratedifferent aspects of the GLASS. The first BS location is nearan Intersection of Grid Squares (IGS), e.g., node 0, and thesecond location is Not near an Intersection of Grid Squares(NIGS), e.g., node 20. The average two-hop neighborhoodsize in this topology is 12.5 sensors. It reflects a denselypopulated network with challenging hidden terminalproblems and multihop data delivery to the BS. Topologies(b), (c), and (d) represent networks with different nodedensities. They include 20, 40, and 80 nodes in a 126� 126 m2

flat area with a BS in a center. We follow the system modeldescribed in Section 3. We assume that each sensor alwayshas pending data ready for transmission, so the simulationsrepresent a data-intensive traffic scenario. For the radiochannel propagation model in the simulations, a two-raypath loss model was chosen and fading was not considered.

The reported simulation results are averaged over 30 runs.Simulations run for 400 seconds of simulation time. Mobilescenarios are discussed later.

We evaluate the network performance using the follow-ing metrics:

Throughput: Amount of the data successfully arrived atthe BS per second.

Success rate: Ratio of the number of transmitted packetsthat reach the BS successfully to the total number ofgenerated packets from the application layer.

Average number of control messages per sensor: Number oftransmitted control messages per sensor per cycle of timeslot generation.

Variation ratio of the packet loss rate: A measure of dispersionbetween the packet loss rate at the 100th second and thepacket loss rates afterward. The network topology has nochanges until 100 seconds. Then, it begins to periodically havea certain node move to a new position in the rest of simulationtime (see details of the mobility scenario in Section 5.4).Packet loss rate ¼ ð1� Success RateÞ.

Average number of cumulative overhead messages per sensor:Number of transmitted control messages per sensor at theend of all time slot recoveries. The time slot recovery isdesigned to update a time slot schedule for all sensors.

Probability of conflict-free time slot schedule in the new two-hop neighborhood: Probability that a sensor’s time slot doesnot conflict with those of its neighbors within a two-hopdistance after the sensor moves to a new location.

Average delay of successful packets: The average time for apacket to successfully reach the BS.

5.1 Transmission Efficiency

To show transmission efficiency at different levels ofnetwork loads, we vary the number of simultaneoussenders from 1 (low contention) to 20 (high contention) inthe random dense network. These simultaneous senders areselected around the BS for each run. Furthermore, weconfigure all sensors in the network as simultaneoussenders and program them to deliver CBR packets fordifferent data generation rates, i.e., (packet/second/node).

We utilize two metrics to demonstrate the transmissionefficiency: throughput and success rate (Fig. 9). We noticethat transmission efficiency of the CSMA is unacceptable forhigh contention scenarios. On the other hand, transmissionefficiency of the DRAND is satisfactory for all traffic loads.In Figs. 9a and 9b, the GLASS protocol is not equipped withthe CAIG function. It shows lower performance compared tothe DRAND protocol, but significantly outperforms theCSMA. Since collisions still occur near intersections of gridcells, the DRAND protocol is better than the GLASS protocolwithout the CAIG function. Overall, the performance withNIGS is higher compared to the case of IGS since, with NIGS,the BS is not near the high-contention area. The success rateof GLASS-IGS with 8-20 senders has a concave shape: moredata are lost with more than 15 senders. Next, the GLASSprotocol with the CAIG function is tested under stringentconditions: 20 simultaneous senders and different datageneration rates. Fig. 9c shows that the GLASS and DRANDare comparable in the success rate metric, while the CSMAperforms worse because of high channel contention andretransmission. Accordingly, it is apparent that the CAIG


Fig. 8. Network Topologies: (a) random dense network, (b) sparsenetwork, (c) medium sparse network, (d) large dense network.

function indeed maintains collision resolution around

intersections of grid cells, and the complete GLASS protocol

matches the DRAND protocol in transmission efficiency

under different traffic loads.Tuning the � parameter can influence the trade-off

between collision avoidance and packet delay. Fig. 9d shows

the collision effect, while the delay effect is demonstrated in

the next results. We use the percentage of sensors without

conflicting time slots as a metric of collision avoidance. In

the random dense network, we observe that collision

avoidance improves with an increase in �. The optimal

value of � depends on network topology. For the topology in

Fig. 8a, an optimal � from simulation is 3.

5.2 Scalability

Next, we here compare performances of the considered

schemes in the large dense network (80 nodes). Fig. 10a

displays the success rate of packets. Both GLASS and

DRAND are consistently reliable because of their conflict-

free time slots assignment. This is in contrast to CSMA,

which demonstrates even lower success rate compared to

the random dense network (see Fig. 9c), since additional

multihop transmissions increase the probability of conflict-

ing transmissions in the large dense network. Fig. 10b

shows the throughput of the CSMA is much lower than thatof the DRAND and GLASS protocols. This is because theCSMA is less efficient in the network with higher channelcontention. Both the DRAND and GLASS protocols aredesigned to overcome challenges of data-intensive trafficscenarios, and so, their performance is better. DRAND andGLASS results are comparable under low traffic loads, e.g.,1 packet/second/node. As the traffic load increases, theDRAND protocol shows higher throughput compared tothe GLASS protocol since its Transmission Frame (TF) isshorter. Both techniques face a bottleneck at the BS: theirthroughputs remain flat at higher traffic loads. Fig. 10cdisplays the latency performance of CSMA, GLASS, andDRAND. The average delay of the successful CSMApackets matches those of the DRAND and GLASS packetsat lower traffic loads (e.g., 1 packet/second/node) since theCSMA’s retransmission function increases the likelihood ofdelivering packets to destinations, if the traffic load is low.As the loads increase and create more packet interference,the CSMA delay performance degrades significantly withhigh variations, while the DRAND and GLASS delayperformances show a smoothly linear increase. Thisphenomenon can be explained by the fact that while thescheduling schemes (e.g., DRAND and GLASS) reduce


Fig. 9. Transmission efficiency in random dense network topology: (a) effect of CAIG on throughput performance, (b) effect of CAIG on packetsuccess rate performance, (c) transmission reliability comparison given � is 3, (d) effect of � on collision avoidance.

Fig. 10. Performances in large dense network topology: (a) packet success rate performance comparison, (b) throughput performance comparison,(c) packet delay performance comparison.

transmission collisions via efficient local coordination, theCSMA results in increasing packet delay (i.e., unreliabletransmission) due to the hidden terminal problem and thehigh channel contention. In addition, because of the shorterlength of DRAND’s TF, the average delay of a successfulDRAND packet is shorter than that of a successful GLASSpacket. We previously mentioned that the � parameterinfluences the length of TF. With a longer TF, averagepacket delay increases, as shown in Fig. 10c. Properselection of � would tune the observed trade-off.

5.3 Overhead Evaluation

Some access control mechanisms may exchange controlpackets to establish channel access. Such control packetshold no application data and consume resources (e.g.,bandwidth, energy, and time), so we regard these packetsas overhead. For scheduling-based approaches, like GLASSand DRAND, the control overhead occurs in both ofNeighborhood Discovery phase (ND) and Time SlotsSchedule and Dissemination phase (TSSD). For CSMA, nocontrol message is used since the RTS/CTS mechanism isnot applied. For a fair comparison, both the GLASS andDRAND protocols adopt the broadcast method to exchangethe control messages.

Fig. 11a shows comparative overhead costs in the randomdense network. DRAND has higher overhead than GLASSdue to extensive messages exchanged for the time slotassignment. GLASS, relying on local sensor computationsoutperforms DRAND in the 3-to-1 ratio. We also explored theeffect of network density on the control overhead. Weperformed experiments with different network topologiesvarying the average number of two-hop neighborhoods

from 0.5 to 11 nodes. The neighborhood size of the networkwas changed by changing the numbers of nodes from 20 to160 within a 150� 150 m2 flat area. Fig. 11b shows thatcontrol overhead with DRAND increases proportionally asthe network density increases. The overhead with GLASS isalmost independent of the network scale and node density. Inaddition, we note that the overhead cost, contributed by theCAIG is low since the CAIG function only involves locallyconflicting sensors. Next, we investigate the convergencetime of executing GLASS for all sensors in the network. Weassume that sensors are not moving during the GLASSprocess. Each sensor starts from performing Grid Searchingand Transmission Frame Assignment to define its spatial andtemporal bounds. After that, sensors run neighborhooddiscovery within their grid cells for 30 seconds to configurean LSM for time slots assignment. Note that the network witha lower node density commonly requires less time to collectneighbor information. In our work, we configure a longperiod, 30 seconds, for the neighborhood discovery takinginto account the networks with different node densities.Finally, sensors perform CAIG to ensure accuracy of theassigned time slots. The runtimes of the above steps for thenetwork topologies in Figs. 8b, 8c, and 8d are approximately32, 32, and 34 seconds, respectively. Such close results areexplained by the low overhead of GLASS protocol.

We performed experiments with networks scaling from20 to 80 nodes, and with node densities from 0.5 to 12nodes per two-hop neighborhood. We observed that thenetwork scale and node density do not have a significantimpact on reliability and overhead cost of GLASS, whichproves its feasibility.

5.4 Impact of Sensor Mobility

In mobile WSNs, the topology may change because ofsensors relocation. Sensor mobility may result in degradingtransmission efficiency and/or increasing overhead cost.Here, we investigate the impact of changing networktopology on GLASS compared to DRAND.

First, we conduct three sets of experiments in the randomdense network with different topology update frequenciesresulting in time slot recovery, i.e., 1, 2, and 4 times. Werandomly select five mobile nodes moving to randomlocations. The node relocations and time slot recovery occurevenly within the simulation duration. We use the Variationratio of the packet loss rate metrics to demonstrate 1) the effectof changed topology on transmission efficiency and 2) theimprovement after time slot recovery. Fig. 12a illustrates thatthe topology changes result in an increase in packet loss untilthe time slot recovery. For instance, GLASS and DRAND with1 TS recovery (red lines) show continuous increases of themetric up to 375 seconds when the time slot recovery occurs.We observe a step trend in the variation ratio increase sincesome node relocations do not impact network performance.For example, with GLASS, the performance does not changemuch from 300 to 350 seconds of simulation time. Morefrequent time slot recovery reduces variations in the packetloss rate and enhances transmission efficiency. Thus, thedegrading influence of the changed topology is alleviated (asshown with the blue and green lines). In other words, aregular update of time slots helps sensors to overcome theimpact of changed topology. The overhead cost caused by the


Fig. 11. Comparison of control overhead: (a) overhead cost in randomdense network, (b) effect of node density on overhead.

time slot recoveries, however, results in extra controlmessages consuming energy and bandwidth [11]. Fig. 12bdisplays this trade-off for GLASS and DRAND. As weobserve, GLASS considerably outperforms DRAND (three-fold decrease in the overhead cost). Note that some time slotrecovery may not enhance network utility if the network isnot changing frequently. Meanwhile, the overhead cost willbe added anyway.

Fig. 12c shows the robustness of the considered schemeswith respect to the changing topology in different networks.Here, we assume that the time slot is not updated afterrelocation of a node. We set up an experiment thatrandomly selects a node with a random location in thenetwork. After that, we check if the current time slot of thenode is conflicting with the time slots of its new two-hopneighbors after the relocation. If it does, this node is in apossible collision area and its time slot schedule is notconflict-free. This is repeated 10 times (i.e., we consider 10randomly moving nodes) in each experiment. Ten trials arerun for each network topology, i.e., Figs. 8b, 8c and 8d, with95 percent confidence intervals represented by error bars.To capture variation of results, we calculate the percentageof similarity between the average results shown by a rednumber above the DRAND column in Fig. 12c. In the sparsenetwork, DRAND and GLASS demonstrate comparableperformance with 92 percent similarity and with theirprobabilities of conflict-free time slot schedule after reloca-tion both being around 0.7. As more nodes join the network(resulting in medium sparse and large dense networks), weobserve that performance of both of the techniques degrade,but degradation of the DRAND protocol is more consider-able compared to the GLASS protocol (only 39 percentsimilarity in the large dense network). This means that inthe large dense network, GLASS better adapts to changingtopology according to the statistical results. This is becauseof the two STFs configuration and the grid concept utilizedin GLASS. In the case of GLASS, there is 50 percent chancethat a node moves to a new location and finds a differenttemporal domain, i.e., a different STF, so potential collisionsare avoided. DRAND protocol does not support such afeature and it is more vulnerable to sensor mobility andchanging network topology.

5.5 Discussion

Sensor applications have different bandwidth requirements.Some WSNs performing telemetry monitoring require lowdata rates while Structural Health Monitoring networksdemand high data throughputs. IEEE 802.15.4 networksperform reasonably well in the case of low data rates. Fig. 9billustrates this fact, which is also confirmed in literature [9],[20]. As the traffic load increases, a more intelligent schemeof GLASS or DRAND is required to handle data-intensivecommunication between sensors. GLASS efficiently miti-gates high data loss in mobile data-intensive sensor net-works. We have demonstrated that GLASS considerablyoutperforms IEEE 802.15.4 contention mode and matchesDRAND in terms of transmission efficiency. Meanwhile,GLASS is much more efficient in adapting to sensor mobilityand changing network topology. At the same time, itprovides a distributed and scalable channel access.

The traffic load has no impact on overhead of the GLASSprotocol regardless of the application requirements. More-over, our experiments demonstrate that the GLASS over-head is practically independent of network scale and nodedensity. In general, the ratio of overhead cost to data cost ishigher in the low data rate network than in the high datarate network. Meanwhile, with the low number of controlmessages for the time slot generation with GLASS (see Fig.11), energy consumption is acceptable even in the case oflow data rates. The overhead generation with DRAND isabout 0.02 percent of the total battery capacity of a nodewith 2,500 mAh and 3 V battery [4]. Apparently, theoverhead cost with GLASS is less than 0.02 percent giventhe same battery. The light weight of GLASS protocolconsiderably enhances its adaptability in the networks withdifferent traffic loads.

GLASS performs channel access using sensor locationinformation. In general, the precision and cost of locationdetermination techniques may be an issue. Recently, this hasbeen partially resolved via the advancement of intelligentand inexpensive positioning techniques [7], [8]. For example,the authors in [7] proposed a distributed location algorithmfor sensors, which is accurate up to 10 centimeters. Mean-while, the average node energy spent during localization atthe networks ranging from 100 to 700 nodes is less than 1 J


Fig. 12. Influence of the changed topology between GLASS and DRAND: (a) transmission efficiency in dynamic networks with time slot recovery,(b) overhead cost of time slot recovery, (c) robustness against changing topology.

(i.e., 0.004 percent of the total battery capacity of a node with2,500 mAh and 3 V battery). That enhances practicability ofthe location-aware channel access, such as the GLASSprotocol. In addition, the GLASS protocol is not sensitiveto the location-based errors since the designs of STF andCAIG function provide temporal and spatial collisionprevention techniques.

The length of TF is an important factor. It affectsbandwidth utilization and packet delay. GLASS andDRAND both adopt the two-hop graph coloring to assigntime slots, and so, the protocols try to minimize the lengthof TF. Because of the two STFs configuration in GLASS, itmay trade small amounts of throughput and packet delayto reduce collisions, control messages overhead, anddegradation due to changing topology, which are importantfor resource-constrained WSNs. Due to the low complexityand small overhead, the cost of time slots recovery inGLASS is considerably lower compared to DRAND.Finally, the GLASS performance is less susceptible tochanging topology.

6 CONCLUSION

In this paper, we presented a novel grid-based schedulingaccess technique called GLASS to mitigate the performancedegradation in data-intensive mobile sensor networks.According to GLASS protocol, each sensor utilizes thetwo-hop graph coloring to derive a collision-free transmis-sion schedule based on the Latin Square Matrix and theCAIG function. The design of a virtual grid networkminimizes the protocol complexity and overhead, andsubsequently improves the scalability. GLASS providesconflict-free time slots assignments in transmission sche-dules, low control messages overhead, and good adapt-ability in dynamic environments.

The results from our analysis and simulations demon-strate the feasibility of our approach in terms of transmis-sion reliability, control overhead, and packet delay. Whileefficiently eliminating conflicting time slots in transmissionschedules, GLASS uses about 33-100 percent of the DRANDoverhead in each generation of time slots scheduling. Theirtransmission reliability is comparable. Our approach isespecially suitable for mobile data-intensive sensor net-works with frequently changing topology.

In future research, we plan to investigate more sophis-ticated cooperation between sensors in the process of timeslots assignment. The design of a virtual grid network ofGLASS is an effective and simple method to cluster localsensors, improving network scalability. However, its opera-tion depends on sensor location awareness. Applying otherclustering techniques (e.g., [43]), we can relax the require-ment of location awareness. Increasing the complexity ofclustering or changing cluster coverage may impact thecontrol messages overhead, which needs further investiga-tion. We also plan to implement and evaluate the GLASSprotocol in a real sensor network testbed.

ACKNOWLEDGMENTS

The authors would like to thank the IEEE Transactions onMobile Computing Associate Editor Prof. Jie Wu and the

anonymous reviewers for their feedback and comments on

this paper. This work is supported by the Telecommunica-

tions and Networking Program at the University of

Pittsburgh and the Q2S Centre at the Norwegian University

of Science and Technology.

REFERENCES

[1] Y. Xu, J. Heidemann, and D. Estrin, “Geography-Informed EnergyConservation for Ad Hoc Routing,” Proc. ACM MobiCom, July2001.

[2] F. Kuhn, R. Wattenhofer, and A. Zollinger, “Worst-Case Optimaland Average-Case Efficient Geometric Ad-Hoc Routing,” Proc.ACM MobiHoc, June 2003.

[3] J.F. Kurose and K.W. Ross, Computer Networking: A Top-DownApproach. Addison-Wesley, 2005.

[4] I. Rhee, A. Warrier, J. Min, and L. Xu, “DRAMD: DistributedRandomized TDMA Scheduling for Wireless Ad Hoc Networks,”Proc. ACM MobiHoc, May 2006.

[5] I. Rhee, A. Warrier, M. Aia, and J. Min, “Z-MAC: A Hybrid MACfor Wireless Sensor Networks,” Proc. Third Int’l Conf. EmbeddedNetworked Sensor Systems (SensSys), Nov. 2005.

[6] Draft Standard IEEE P802.15.4/D18, Low Rate Wireless Personal AreaNetworks, IEEE, Feb. 2003.

[7] A. Savvides, C.-C. Han, and M.B. Strivastava, “Dynamic Fine-Grained Localization in Ad-Hoc Networks of Sensors,” Proc. ACMMobiCom, July 2001.

[8] L. Doherty, K.S.J. Pister, and L.E. Ghaoui, “Convex PositionEstimation in Wireless Sensor Networks,” Proc. IEEE INFOCOM,Apr. 2001.

[9] N. Xu, S. Rangwala, K.K. Chintalapudi, D. Ganesan, A. Broad, R.Govindan, and D. Estrin, “A Wireless Sensor Network forStructural Monitoring,” Proc. Second Int’l Conf. Embedded NetworkedSensor Systems (SenSys), Nov. 2004.

[10] NS-2: The Network Simulator, http://www.isi.edu/nsnam/ns,2006.

[11] C. Siva Ram Murthy and B.S. Manoj, Ad Hoc Wireless Networks—Architectures and Protocols, p. 230. Prentice Hall, 2003.

[12] M.J. Lee and J. Zheng, “Will IEEE 802.15.4 Make UbiquitousNetworking a Reality? A Discussion on a Potential Low Power,Low Bit Rate Standard,” IEEE Comm. Magazine, vol. 42, no. 6,pp. 140-146, June 2004.

[13] C.F. Laywine and G.L. Mullen, Discrete Mathematics Using LatinSquares. John Wiley & Sons, 1998.

[14] J.V. Bradley, “Complete Counterbalancing of Immediate Se-quential Effects in a Latin Square Design,” J. Am. StatisticalAssoc., vol. 53, no. 284, pp. 525-528, Dec. 1958.

[15] G. Zhou, T. He, S. Krishnamurthy, and J.A. StanKovic, “Impact ofRadio Irregularity on Wireless Sensor Networks,” Proc. ACMMobiSys, June 2004.

[16] T. Herman and S.A. Tixeuil, “A Distributed TDMA Slot Assign-ment Algorithm for Wireless Sensor Networks,” Proc. Int’l Work-shop Algorithmic Aspect of Wireless Sensor Networks (AlgoSensors),2004.

[17] S. Madden, M.J. Franklin, J.M. Hellerstein, and W. Hong, “TheDesign of an Acquisitional Query Processor for Sensor Networks,”Proc. ACM SIGMOD, June 2003.

[18] L.F.W. Van Hoesel and P.J.M. Havinga, “A Lightweight MediumAccess Protocol (LMAC) for Wireless Sensor Networks: ReducingPreamble Transmissions and Transceiver State Switches,” Proc.Int’l Conf. Networked Sensing Systems, June 2004.

[19] P. Scerri, Y. Xu, E. Liao, J. Lai, and K. Sycara, “Scaling Teamworkto Very Large Teams,” Proc. Int’l Joint Conf. Autonomous Agents andMultiagent Systems (AAMAS), Aug. 2004.

[20] C.-K. Lin, V. Zadorozhny, and P. Krishnamurthy, “Efficient DataDelivery in Wireless Sensor Networks: Algebraic Cross-LayerOptimization versus CSMA/CA,” Int’l J. Ad Hoc and SensorWireless Networks, vol. 4, nos. 1/2, pp. 149-174, 2007.

[21] C.-K. Lin, V. Zadorozhny, and P. Krishnamurthy, “EfficientHybrid Channel Access for Data Intensive Sensor Networks,”Proc. IEEE Int’l Workshop Heterogeneous Wireless Networks, May2007.

[22] V. Zadorozhny, D. Sharma, P. Krishnamurthy, and A. Labrinidis,“Tuning Query Performance in Sensor Databases,” Proc. Int’l Conf.Mobile Data Management (MDM), May 2005.


[23] C. Bererton, L. Navarro-Serment, R. Grabowski, C. Paredis, and P.Khosla, “Millibots: Small Distributed Robots for Surveillance andMapping,” Proc. Govt. Microcircuit Applications Conf., Mar. 2000.

[24] G. Sibley, M. Rahimi, and G. Sukhatme, “Robomote: A TinyMobile Robot Platform for Large-Scale Ad Hoc Sensor Networks,”Proc. Int’l Conf. Robotics and Automation, May 2002.

[25] S. Bergbreiter and K.S. Pister, “CotsBots: An Off-The-ShelfPlatform for Distributed Robotics,” Proc. IEEE/RSJ Int’l Conf.Intelligent Robots and Systems, 2004.

[26] P. Scerri, D. Pynadath, L. Johnson, P. Rosenbloom, M. Si, N.Schurr, and M. Tambe, “A Prototype Infrastructure for Distrib-uted Robot Agent Person Teams,” Proc. Int’l Joint Conf. Autono-mous Agents and Multiagent Systems (AAMAS), 2003.

[27] H. Zhang, H. Shen, and H. Kan, “Distributed Tuning AttemptProbability for Data Gathering in Random Access Wireless SensorNetworks,” Proc. Int’l Conf. Advanced Information Networking andApplications, Apr. 2006.

[28] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci,“Wireless Sensor Networks: A Survey,” Computer Networks, vol. 38,no. 4, pp. 393-422, 2002.

[29] W. Ye, J. Heidemann, and D. Estrin, “An Energy Efficient MACProtocol for Wireless Sensor Networks,” Proc. IEEE INFOCOM,June 2002.

[30] K.M. Sivalingam, J.-C. Chen, P. Agrawal, and M.B. Srivastava,“Design and Analysis of Low Power Access Protocols for Wirelessand Mobile ATM,” Wireless Network, vol. 6, pp. 73-87, 2000.

[31] S. Chatterjea, L.F.W. van Hoesel, and P.J.M. Havinga, “AI-LMAC:An Adaptive, Information-Centric and Lightweight MAC Protocolfor Wireless Sensor Networks,” Proc. DEST Int’l Workshop SignalProcessing for Sensor Networks, 2004.

[32] M.H. Ammar and D.S. Stevens, “A Distributed TDMA Reschedul-ing Procedure for Mobile Packet Radio Networks,” Proc. IEEEConf. Comm. (ICC), June 1991.

[33] F.N. Ali, P.K. Appani, J.L. Hammond, V.V. Mehta, D.L. Noneaker,and H.B. Russell, “Distributed and Adaptive TDMA Algorithmsfor Multiple-Hop Mobile Networks,” Proc. IEEE Military Comm.Conf. (MILCOM), Oct. 2002.

[34] N. Trigoni, Y. Yao, A. Demers, and J. Gehrke, “Wave Scheduling:Energy-Efficient Data Dissemination for Sensor Networks,” Proc.Int’l Workshop Data Management for Sensor Networks, Aug. 2004.

[35] N. Vaidya and M. Miller, “A MAC Protocol to Reduce SensorNetwork Energy Consumption Using a Wakeup Radio,” IEEETrans. Mobile Computing, vol. 4, no. 3, pp. 228-242, May/June 2005.

[36] C. Schurgers, V. Tsiatsis, S. Ganeriwal, and M. Srivastava,“Topology Management for Sensor Networks: Exploiting Latencyand Density,” Proc. ACM MobiHoc, June 2002.

[37] C.-K. Lin, V. Zadorozhny, and P. Krishnamurthy, “Grid-BasedAccess Scheduling for Mobile Data Intensive Sensor Networks,”Proc. Int’l Conf. Mobile Data Management (MDM), May 2008.

[38] L. Bao, “MALS: Multiple Access Scheduling Based on LatinSquares,” Proc. IEEE Military Comm. Conf., Oct. 2004.

[39] I. Jawhar and J. Wu, “Race-Free Resource Allocation for QoSSupport in Wireless Networks,” Int’l J. Ad Hoc and Sensor WirelessNetworks, vol. 1, no. 3, pp. 1-28, 2005.

[40] J.H. Ju and V.O.K. Li, “TDMA Scheduling Design of MultihopPacket Radio Networks Based on Latin Squares,” IEEE J. SelectedAreas in Comm., vol. 17, no. 8, pp. 1345-1352, Aug. 1999.

[41] I. Gerasimov and R. Simon, “A Bandwidth-Reservation Mechan-ism for On-Demand Ad Hoc Path Finding,” Proc. IEEE 35th Ann.Simulation Symp., Apr. 2002.

[42] D. Zhu, Q. Qi, Y. Wang, K.-M. Lee, and S. Foong, “A PrototypeMobile Wireless Sensor Network for Structural Health Monitor-ing,” Proc. SPIE Conf. Nondestructive Characterization for CompositeMaterials, Aerospace Eng., Civil Infrastructure and Homeland Security,Mar. 2009.

[43] F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang, “A Two-Tier DataDissemination Model for Large-Scale Wireless Sensor Networks,”Proc. ACM MobiCom, Sept. 2002.

Chih-Kuang Lin received the MSc and PhDdegrees in telecommunications and networkingfrom the University of Colorado, Boulder, andthe University of Pittsburgh in 2000 and 2008,respectively. Currently, he is a postdoctoralresearch fellow in the Q2S Centre at theNorwegian University of Science and Technol-ogy. From 2000 to 2003, he was a systemengineer in Samsung Telecom. His researchinterests are in the field of networking and

communication networks with focus on game-theoretic models, queuingtheory and protocols of wireless systems, and MANET and sensornetworks. He is a student member of the IEEE.

Vladimir I. Zadorozhny received the PhDdegree from the Institute for Problems of Infor-matics, Russian Academy of Sciences, Moscow,in 1993. He is an associate professor in theSchool of Information Sciences, University ofPittsburgh. Before coming to United States, hewas a principal research fellow in the Institute ofSystem Programming, Russian Academy ofSciences. Since May 1998, he has been workingas a research associate at the Institute for

Advanced Computer Studies, University of Maryland, College Park. Hejoined the University of Pittsburgh in September 2001, where he is thehead of the Network Aware Data Management Group and a member ofthe Advanced Data Management Technologies Laboratory. His researchinterests include networked information systems, complex adaptivesystems, WSN data management, query optimization in resource-constrained distributed environments, and scalable architectures forwide-area environments with heterogeneous information servers. He is amember of the IEEE.

Prashant V. Krishnamurthy is an associateprofessor in the School of Information Sciences,University of Pittsburgh, where he regularlyteaches courses on wireless communicationsystems and networks, cryptography, and net-work security. His research interests are inwireless network security, wireless data net-works, position location in indoor wireless net-works, and radio channel modeling for indoorwireless networks. He also served as the chair of

the IEEE Communications Society Pittsburgh Chapter from 2000 to2005. He is a member of the IEEE.

Ho-Hyun Park received the BS, MS, and PhDdegrees in computer science and engineeringfrom the Seoul National University and KAIST in1987, 1995, and 2001, respectively. From 1987to 2003, he was a principal engineer at SamsungElectronics. Currently, he is an associate pro-fessor of electrical and electronics engineeringat Chung-Ang University. His research interestsinclude multimedia streaming, retrieval, anddatabases, spatiotemporal databases, LBS,

ubiquitous computing, and real time and embedded systems.

Chan-Gun Lee received the BS, MS, and PhDdegrees in computer science from Chung-AngUniversity, KAIST, and the University of Texas,Austin, in 1996, 1998, and 2005, respectively.From 2005 to 2007, he was a senior softwareengineer at Intel. Currently, he is an assistantprofessor of computer science and engineeringat Chung-Ang University. His research interestsspan real-time systems, software engineeringfor time-critical systems, test frameworks, and

event-based systems.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

A Distributed and Scalable Time Slot Allocation Protocol for Wireless Sensor Networks_Mobile Computing_2011